On Fri, Jun 5, 2009 at 9:15 AM, Agblad Tore<tore.agb...@volvo.com> wrote:

> Thanks' for the info Rob and Hans-Joachim !
> But I read last in this email that performance get worse.
> I guess that does not depend on the cpuplugd itself is running, but because
> the management of memory uses more cpu ? Am I right ?
> Is it the case the that for activating CPUs it's ok to use but the 
> memplug/unplug
> does not give any gain ?
> I'm just now reading the pdfs, but they are big, will take some time.
> Probably the answer is there ;)

The answer is probably not there yet ;-) Sorry to disappoint you.

I did look into this a bit. I expected the z/VM API to be the reason
(because the CMM driver does a CP DIAG call for each page to release)
but that turns out to be pretty cheap. So it's probably Linux walking
the memory management tables to find out which pages to free up.
The other factor is the extra cost for excessive swapping (to VDISK
hopefully) when the workload has started and memplug is slowly adding
memory. This process is slow because of the static increment. Because
cpuplugd can't tell how much excessive memory you have, it uses a
"kaasschaaf" approach and it takes a long time for memory to get
trimmed (or you would choke the penguin with the first stroke).
Similar, adding memory is done little by little to observe the effect.
It's like fighting a fire with cups of tea ;-)  Given enough beer, I
might be able to come up with interesting ideas on how to tweak it
(using different metrics) but I really don't believe in doing this
without z/VM data.

Using cpuplugd for CPU management is another discussion (and some
slides in my "CPU" presentation at the same place). There's one
particular lab case where I could demonstrate a benefit, but I fear
that one may not be a common real life scenario. And there's a design
issue in z/VM 5.4 that has some implications if you use this. Others
have reported the resource usage of cpuplugd itself (when reading the
metrics every 10 seconds or so) is a concern.
My recommendation would be to stay away from it. Proper planning and
performance evaluation is going to give you better results.

Rob
--
Rob van der Heij
Velocity Software
http://www.velocitysoftware.com/

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to lists...@vm.marist.edu with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

Reply via email to